Optimasi Model XGBoost Dengan Seleksi Fitur Mutual Information Dan Threshold Tuning Untuk Deteksi Intrusi Jaringan

Abstract

Network Intrusion Detection Systems (NIDS) are essential for protecting networks from evolving cyber threats. Although Machine Learning can be relied upon for NIDS, challenges remain in achieving high accuracy while maintaining low false alarm rates. This research proposes an optimized NIDS framework using the Extreme Gradient Boosting (XGBoost) algorithm, which is enhanced through systematic feature selection and hyperparameter tuning. The methodology integrates a two-stage feature selection process that combines ExtraTreesClassifier for initial importance analysis and SelectKBest with mutual information for identifying the optimal feature subset. Hyperparameter optimization is performed using RandomizedSearchCV with 5-fold cross-validation, followed by threshold calibration to balance the False Positive Rate (FPR) and False Negative Rate (FNR). The model is trained and evaluated on the UNSW-NB15 dataset, which contains 257,673 network traffic records with binary classification (normal vs. attack). The results of the experiment show that the optimized XGBoost model achieved an accuracy of 95.4%, precision of 94.81%, recall of 95.29%, F1-score of 95.04%, and a significantly reduced FPR of 5.09%. The feature selection process identified 37 most informative features from the original 42 features, which contributed to improved model performance and efficiency. These findings indicate that an integrated approach of adaptive feature selection and systematic model optimization effectively improves intrusion detection performance, offering a robust and balanced solution for modern network security applications.